Barry Kelly
@barrkel.bsky.social
Software engineer and architect. Zurich, London, Galway. Google, Duco, Delphi.
Reposted by Barry Kelly
2/6
This is the point I have been making again and again over the years. The global economy is a closed system, and it must balance. This means that domestic imbalances created by countries that control their external accounts must...
This is the point I have been making again and again over the years. The global economy is a closed system, and it must balance. This means that domestic imbalances created by countries that control their external accounts must...
December 8, 2025 at 4:30 AM
2/6
This is the point I have been making again and again over the years. The global economy is a closed system, and it must balance. This means that domestic imbalances created by countries that control their external accounts must...
This is the point I have been making again and again over the years. The global economy is a closed system, and it must balance. This means that domestic imbalances created by countries that control their external accounts must...
There's a problem with LLM problem solving that I'm seeing pretty consistently now.
LLMs, when they encounter an error, formulate a hypothesis about what's wrong.
Then, rather than validate the hypothesis, they immediately try to implement a solution based on the hypothesis.
Check the hypothesis!
LLMs, when they encounter an error, formulate a hypothesis about what's wrong.
Then, rather than validate the hypothesis, they immediately try to implement a solution based on the hypothesis.
Check the hypothesis!
November 28, 2025 at 9:38 PM
There's a problem with LLM problem solving that I'm seeing pretty consistently now.
LLMs, when they encounter an error, formulate a hypothesis about what's wrong.
Then, rather than validate the hypothesis, they immediately try to implement a solution based on the hypothesis.
Check the hypothesis!
LLMs, when they encounter an error, formulate a hypothesis about what's wrong.
Then, rather than validate the hypothesis, they immediately try to implement a solution based on the hypothesis.
Check the hypothesis!
This isn't really accurate but understanding why is quite subtle.
Mimicking patterns of words relies on approximate knowledge. LLMs have approximate knowledge of lots of things.
This means they're usually right on the gist of something but wrong on specifics. 1/
Mimicking patterns of words relies on approximate knowledge. LLMs have approximate knowledge of lots of things.
This means they're usually right on the gist of something but wrong on specifics. 1/
Chatbots — LLMs — do not know facts and are not designed to be able to accurately answer factual questions. They are designed to find and mimic patterns of words, probabilistically. When they’re “right” it’s because correct things are often written down, so those patterns are frequent. That’s all.
October 27, 2025 at 6:48 PM
This isn't really accurate but understanding why is quite subtle.
Mimicking patterns of words relies on approximate knowledge. LLMs have approximate knowledge of lots of things.
This means they're usually right on the gist of something but wrong on specifics. 1/
Mimicking patterns of words relies on approximate knowledge. LLMs have approximate knowledge of lots of things.
This means they're usually right on the gist of something but wrong on specifics. 1/
Reposted by Barry Kelly
the most exciting specialization method, imo, is Cartridges
basically take a gigantic context, and then *train* a KV cache
this is wild, who even thinks to train a KV cache? but it works. incredible compression & recall
hazyresearch.stanford.edu/blog/2025-06...
basically take a gigantic context, and then *train* a KV cache
this is wild, who even thinks to train a KV cache? but it works. incredible compression & recall
hazyresearch.stanford.edu/blog/2025-06...
October 15, 2025 at 11:39 AM
the most exciting specialization method, imo, is Cartridges
basically take a gigantic context, and then *train* a KV cache
this is wild, who even thinks to train a KV cache? but it works. incredible compression & recall
hazyresearch.stanford.edu/blog/2025-06...
basically take a gigantic context, and then *train* a KV cache
this is wild, who even thinks to train a KV cache? but it works. incredible compression & recall
hazyresearch.stanford.edu/blog/2025-06...
Reposted by Barry Kelly
Nathan Gill, Reform's leader in Wales at the last Welsh parliament election, has pled guilty to 8 bribery charges, taking Russian cash while he was one of Nigel Farage’s MEPs to speak in favour of Putin's Russia in that parliament
www.theguardian.com/uk-news/2025...
www.theguardian.com/uk-news/2025...
Reform UK’s ex-leader in Wales Nathan Gill pleads guilty to bribery charges
Gill admits to eight counts of bribery relating to pro-Russia statements made in the European parliament and articles
www.theguardian.com
September 26, 2025 at 3:03 PM
Nathan Gill, Reform's leader in Wales at the last Welsh parliament election, has pled guilty to 8 bribery charges, taking Russian cash while he was one of Nigel Farage’s MEPs to speak in favour of Putin's Russia in that parliament
www.theguardian.com/uk-news/2025...
www.theguardian.com/uk-news/2025...
Reposted by Barry Kelly
The "Kirk was civil" lady from the Atlantic has been stuck in my head this morning. So much of this is about existing elites focusing on maintaining elite status in the new order. The Zyuganovication of American liberalism.
September 15, 2025 at 8:31 AM
The "Kirk was civil" lady from the Atlantic has been stuck in my head this morning. So much of this is about existing elites focusing on maintaining elite status in the new order. The Zyuganovication of American liberalism.
I'm kind of disappointed in GPT5.
I had a tough problem recently. zfsbootmenu handoff to Ubuntu, failure for framebuffer to initialize, making it harder to configure my new rig.
4-5 minute waits for advice that wasn't that useful.
Gemini Pro 2.5 was significantly more useful, and faster.
I had a tough problem recently. zfsbootmenu handoff to Ubuntu, failure for framebuffer to initialize, making it harder to configure my new rig.
4-5 minute waits for advice that wasn't that useful.
Gemini Pro 2.5 was significantly more useful, and faster.
September 5, 2025 at 9:06 PM
I'm kind of disappointed in GPT5.
I had a tough problem recently. zfsbootmenu handoff to Ubuntu, failure for framebuffer to initialize, making it harder to configure my new rig.
4-5 minute waits for advice that wasn't that useful.
Gemini Pro 2.5 was significantly more useful, and faster.
I had a tough problem recently. zfsbootmenu handoff to Ubuntu, failure for framebuffer to initialize, making it harder to configure my new rig.
4-5 minute waits for advice that wasn't that useful.
Gemini Pro 2.5 was significantly more useful, and faster.
Reposted by Barry Kelly
a state that cannot enforce consequences for a coup attempt is a state that is going to fail, & so it turned out to be
August 30, 2025 at 7:11 PM
a state that cannot enforce consequences for a coup attempt is a state that is going to fail, & so it turned out to be
Reposted by Barry Kelly
'Immigration doesn't bother me but it seems to bother everyone else' – Everyone.
This is what happens when the media runs away with Hard Right narratives.
This is what happens when the media runs away with Hard Right narratives.
August 26, 2025 at 5:37 AM
'Immigration doesn't bother me but it seems to bother everyone else' – Everyone.
This is what happens when the media runs away with Hard Right narratives.
This is what happens when the media runs away with Hard Right narratives.
Imagine if we got this 20 years ago, something able to understand and write language, that fits on less than half a CD, but requires monstrous amounts of computation.
It would be like alien technology, practically beyond our comprehension.
It would be like alien technology, practically beyond our comprehension.
The new Gemma 3 270M open weights model from Google is really fun - it's absolutely tiny, just a 241MB download
I asked it for an SVG of a pelican riding a bicycle and it wrote me a delightful little poem instead
simonwillison.net/2025/Aug/14/...
I asked it for an SVG of a pelican riding a bicycle and it wrote me a delightful little poem instead
simonwillison.net/2025/Aug/14/...
August 15, 2025 at 11:11 AM
Imagine if we got this 20 years ago, something able to understand and write language, that fits on less than half a CD, but requires monstrous amounts of computation.
It would be like alien technology, practically beyond our comprehension.
It would be like alien technology, practically beyond our comprehension.
Reposted by Barry Kelly
Misinformation = false or misleading information spread by people who believe it’s true.
Malinformation = true information used in misleading or harmful ways.
Malinformation = true information used in misleading or harmful ways.
June 27, 2025 at 6:53 AM
Misinformation = false or misleading information spread by people who believe it’s true.
Malinformation = true information used in misleading or harmful ways.
Malinformation = true information used in misleading or harmful ways.
Reposted by Barry Kelly
Media scholar Clare Wardle and colleagues developed a helpful typology:
Disinformation = false or misleading information spread deliberately to deceive or manipulate.
Disinformation = false or misleading information spread deliberately to deceive or manipulate.
June 27, 2025 at 6:53 AM
Media scholar Clare Wardle and colleagues developed a helpful typology:
Disinformation = false or misleading information spread deliberately to deceive or manipulate.
Disinformation = false or misleading information spread deliberately to deceive or manipulate.
This is super interesting stuff. Adds quite a bit of nuance on top of the simplistic Economics 101 model of thinking scarcity means it's fine to let prices rise. It can actually be an opportunity to collude with competitors and raise prices!
Our qualitative analysis confirms implicit pricing coordination through cost shocks. Example: "We're very aware of what's going on in the industry, and you're seeing some fairly aggressive moves on pricing. I think there is an ability to price and permission to price." 7/9
June 17, 2025 at 5:00 PM
This is super interesting stuff. Adds quite a bit of nuance on top of the simplistic Economics 101 model of thinking scarcity means it's fine to let prices rise. It can actually be an opportunity to collude with competitors and raise prices!
There's something I'm starting to get sick of in LLM prose. It's not flattery - it's antithesis. It's starting out not with the point, but instead its opposite. The forced contrast is not effective, it's overused. And I don't think I'm wrong to see it as a tell.
June 9, 2025 at 9:50 AM
There's something I'm starting to get sick of in LLM prose. It's not flattery - it's antithesis. It's starting out not with the point, but instead its opposite. The forced contrast is not effective, it's overused. And I don't think I'm wrong to see it as a tell.
Reposted by Barry Kelly
I just had a vision of that Oval Office ambush when Tr*mp accused Zelenskyy of having "no cards". Zelenskyy must have sat there smiling inside, thinking, "You've no fucking idea what cards I've got you dumb moron".
June 1, 2025 at 6:46 PM
I just had a vision of that Oval Office ambush when Tr*mp accused Zelenskyy of having "no cards". Zelenskyy must have sat there smiling inside, thinking, "You've no fucking idea what cards I've got you dumb moron".
Reposted by Barry Kelly
Announcing AlphaEvolve, our new LLM coding agent that has
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: deepmind.google/discover/blo...
White paper:
storage.googleapis.com/deepmind-med...
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: deepmind.google/discover/blo...
White paper:
storage.googleapis.com/deepmind-med...
AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms
New AI agent evolves algorithms for math and practical applications in computing by combining the creativity of large language models with automated evaluators
deepmind.google
May 14, 2025 at 8:11 PM
Announcing AlphaEvolve, our new LLM coding agent that has
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: deepmind.google/discover/blo...
White paper:
storage.googleapis.com/deepmind-med...
- made new scientific discoveries
- discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers)
Blog: deepmind.google/discover/blo...
White paper:
storage.googleapis.com/deepmind-med...
Reposted by Barry Kelly
Thanks to this anonymous user, I now have preferred terms for output tokens and thinking tokens: “yap” and “juice.” Future model cards should include yap & juice statistics.
Now we just need as good a term for the size of the context window.
Now we just need as good a term for the size of the context window.
Is this on a plus or pro plan? I’ve heard that yap and juice are adjusted dynamically based on which plan you have so it can spend more cycles thinking.
May 6, 2025 at 2:29 PM
Thanks to this anonymous user, I now have preferred terms for output tokens and thinking tokens: “yap” and “juice.” Future model cards should include yap & juice statistics.
Now we just need as good a term for the size of the context window.
Now we just need as good a term for the size of the context window.
You can't out-Nazi the Nazis. Awareness of how the dynamic plays out in practice isn't widespread, I think.
Given the local election results in the UK, a short #thread - based on much of our own research - on the vicious cycle of radical right support that we have seen in many countries.
1) When radical right parties (such as Reform) are electorally successful established parties move right on immigration
1) When radical right parties (such as Reform) are electorally successful established parties move right on immigration
May 2, 2025 at 10:22 AM
You can't out-Nazi the Nazis. Awareness of how the dynamic plays out in practice isn't widespread, I think.
Reposted by Barry Kelly
So the Supreme Court’s strategy seems pretty clearly to be forestall admitting we have a constitutional crisis by giving Trump huge victories on narrow procedural grounds. This is akin to having engineers move the Rubicon River closer and closer to Rome to avoid Caesar crossing it.
April 8, 2025 at 7:07 PM
So the Supreme Court’s strategy seems pretty clearly to be forestall admitting we have a constitutional crisis by giving Trump huge victories on narrow procedural grounds. This is akin to having engineers move the Rubicon River closer and closer to Rome to avoid Caesar crossing it.
Reposted by Barry Kelly
1/2
Amid all the huffing and puffing, Brown University's Mark Blyth takes a longer view: "Ending the current system will be massively disruptive, no doubt, and the prospect of...
prosyn.org/rFWknNO
Amid all the huffing and puffing, Brown University's Mark Blyth takes a longer view: "Ending the current system will be massively disruptive, no doubt, and the prospect of...
prosyn.org/rFWknNO
Trump, Tariffs, and the Fate of the Dollar | by Mark Blyth - Project Syndicate
Mark Blyth argues that America’s protectionist policies reflect a global economic reordering that was already underway.
prosyn.org
April 4, 2025 at 1:12 PM
1/2
Amid all the huffing and puffing, Brown University's Mark Blyth takes a longer view: "Ending the current system will be massively disruptive, no doubt, and the prospect of...
prosyn.org/rFWknNO
Amid all the huffing and puffing, Brown University's Mark Blyth takes a longer view: "Ending the current system will be massively disruptive, no doubt, and the prospect of...
prosyn.org/rFWknNO
Reposted by Barry Kelly
Putin's not anti-immigration!
It was the thing that made him unpopular among even the real ethno-nats and why Libs like Navalny picked it up because how salient it was. Immigration from CA is massive and the border is effectively "open"
It was the thing that made him unpopular among even the real ethno-nats and why Libs like Navalny picked it up because how salient it was. Immigration from CA is massive and the border is effectively "open"
The right in Western Europe and the US has long admired how Putin governs: the strongman style, the intense opposition to LGBT+ "ideology", the anti-immigrant sentiment, the disregard for the rule of law, etc. This ideological affinity is a big reason why the US wants to cut loose Ukraine.
February 14, 2025 at 10:23 AM
Putin's not anti-immigration!
It was the thing that made him unpopular among even the real ethno-nats and why Libs like Navalny picked it up because how salient it was. Immigration from CA is massive and the border is effectively "open"
It was the thing that made him unpopular among even the real ethno-nats and why Libs like Navalny picked it up because how salient it was. Immigration from CA is massive and the border is effectively "open"
Reposted by Barry Kelly
Mom would have been 82 today. She died at 55. I recently realized that I am now older than she ever was, which seems like a fundamental wrongness in the universe.
I was particularly thinking about her penchant for taking a direct and vivid role in her schools’ disciplinary approaches. /1
I was particularly thinking about her penchant for taking a direct and vivid role in her schools’ disciplinary approaches. /1
January 14, 2025 at 7:42 PM
Mom would have been 82 today. She died at 55. I recently realized that I am now older than she ever was, which seems like a fundamental wrongness in the universe.
I was particularly thinking about her penchant for taking a direct and vivid role in her schools’ disciplinary approaches. /1
I was particularly thinking about her penchant for taking a direct and vivid role in her schools’ disciplinary approaches. /1
I want two dimensional restaurant ratings. One for how refined the food and experience is, the other for how good it is within this dimension.
You don't go to the best pizza place for fine dining. But pizza is easy. Sit down dining with a wine menu can be hard to pin down from reviews alone.
You don't go to the best pizza place for fine dining. But pizza is easy. Sit down dining with a wine menu can be hard to pin down from reviews alone.
December 30, 2024 at 2:09 PM
I want two dimensional restaurant ratings. One for how refined the food and experience is, the other for how good it is within this dimension.
You don't go to the best pizza place for fine dining. But pizza is easy. Sit down dining with a wine menu can be hard to pin down from reviews alone.
You don't go to the best pizza place for fine dining. But pizza is easy. Sit down dining with a wine menu can be hard to pin down from reviews alone.
When my son was a toddler, he got used to the Alexa wakeword being "computer" (a la Star Trek). So now he thinks computers, laptops and tablets are three distinct classes of computing device, and computers are the things you talk to.
December 30, 2024 at 2:05 PM
When my son was a toddler, he got used to the Alexa wakeword being "computer" (a la Star Trek). So now he thinks computers, laptops and tablets are three distinct classes of computing device, and computers are the things you talk to.
Reposted by Barry Kelly
(For what it's worth, which isn't a lot, Clause agrees I'm entitled to feel offended.)
December 18, 2024 at 9:20 AM
(For what it's worth, which isn't a lot, Clause agrees I'm entitled to feel offended.)